TWI744798B - Evaluation method and system of neuropsychiatric diseases based on brain imaging - Google Patents

Abstract

A neuropsychiatric disease assessment method based on brain imaging includes: sequentially performing azimuth correction processing and brain tissue recognition processing on the 3D brain images of multiple normal persons and the 3D brain images of multiple neuropsychiatric patients, and obtaining each A white matter, gray matter, and spinal fluid image parts of a 3D brain image; according to all white matter image parts, all gray matter image parts, and all spinal fluid image parts processed by dimensionality reduction, the deep neural network algorithm is used for self-learning and training to obtain the corresponding The first to third judgment models of gray matter, white matter, and spinal fluid; and the white matter, gray matter, and spinal fluid images of the 3D brain images to be judged are fed into the first to third judgment models after dimensionality reduction. After calculation, the first to third judgment results are obtained respectively, and the evaluation results are generated accordingly.

Description

Evaluation method and system of neuropsychiatric diseases based on brain imaging

The present invention relates to the assessment of neuropsychiatric diseases, in particular to a neuropsychiatric assessment method and system based on brain images.

With the pressure of competition brought about by the rapid economic development, the patients suffering from mental illnesses such as schizophrenia (also known as schizophrenia), bipolar disorder (Bipolar), and depression (Depression) in countries around the world have already increase yearly. In addition, the aging population in the world has also increased the number of people suffering from neurodegenerative diseases such as Dementia and Parkinson’s Disease. At present, these neuropsychiatric diseases are still diagnosed by qualitative analysis based on observation records of patient behavior and the clinical experience of neuropsychiatrists. However, since the cognitive or behavioral manifestations of a variety of neuropsychiatric diseases have different degrees of overlap in their representations, diagnosis based solely on the symptoms or behavioral manifestations of the patients will be significantly insufficient. Therefore, clinically, a non-invasive imaging technique, such as MRI brain images obtained by a Magnetic Resonance Imaging (MRI) instrument, can be used to further understand and understand the complexity and variability of neuropsychiatric diseases.

Therefore, how to use brain imaging to evaluate neuropsychiatric diseases as a clinical diagnostic aid for psychiatrists has become an important research topic.

Therefore, an object of the present invention is to provide a neuropsychiatric disease assessment method based on brain imaging, which can overcome the above-mentioned problems or at least partially solve the above-mentioned problems.

Therefore, the method for evaluating neuropsychiatric diseases based on brain imaging provided by the present invention is executed by a computer system and includes the following steps: (A) Obtain a plurality of 3D images corresponding to a plurality of normal persons without neuropsychiatric diseases. Brain images, and multiple 3D brain images corresponding to multiple patients suffering from neuropsychiatric diseases; (B) For each 3D brain image obtained in step (A), perform azimuth correction processing, one of which is deep The learning algorithm is used to identify the first to third 2D image parts that are captured from the 3D brain image and are orthogonal to each other and intersect at the center of the 3D brain image, so as to identify the first to third 2D image parts Results Correct the 3D azimuth angle of the center of the 3D brain image; (C) For each corrected 3D brain image, perform a brain tissue recognition process based on the different voxel characteristics of white matter, gray matter and spinal fluid on the brain image In order to extract a white matter image portion, a gray matter image portion, and a spinal fluid image portion from the 3D brain image; (D) all white matter image portions and all gray matter images obtained according to step (C) and subjected to dimensionality reduction processing The imaging part and all spinal fluid imaging parts, using the depth of God Through self-learning and training of network algorithms, first to third judgment models corresponding to gray matter, white matter, and spinal fluid are obtained. Each of the first to third judgment models includes an input layer, one and a neuropsychiatric patient Feature selection layer, multiple hidden layers and an output layer related to abnormal brain regions; (E) When a 3D brain image to be determined is received, for the 3D brain image to be determined, the azimuth angle is executed sequentially Correction processing and the brain tissue recognition processing to obtain a white matter image portion to be determined, a gray matter image portion to be determined, and a spinal fluid image portion to be determined corresponding to the 3D brain image to be determined; (F) will undergo dimensionality reduction processing The subsequent white matter image portion to be determined, the gray matter image portion to be determined, and the spinal fluid image portion to be determined are fed into the input layer of the first to third determination models, respectively, through the features of the first to third determination models After the calculation of the selected layer and the hidden layers, the first to third judgment results are generated in the output layers of the first to third judgment models, respectively; and (G) a related result is generated according to the first to third judgment results Evaluation results of abnormal brain areas.

Therefore, another object of the present invention is to provide a neuropsychiatric disease assessment system based on brain imaging, which overcomes the above-mentioned problems or at least partially solves the above-mentioned problems.

Therefore, the neuropsychiatric disease assessment system based on brain imaging provided by the present invention includes a judgment platform. The judgment platform includes a storage module, a transmission module, an azimuth correction module, a brain tissue recognition module, and a processing module. The storage module stores the first to second Three judgment models. The transmission module is suitable for receiving a 3D brain image to be determined. The azimuth correction module is connected to the transmission module to receive the 3D brain image to be determined, and from the 3D brain image to be determined, first to third 2D orthogonal to each other and intersecting at the center of the 3D brain image to be determined are extracted The image part uses a deep learning algorithm to perform image recognition of the first to third 2D image parts, and corrects the 3D of the center of the to-be-determined 3D brain image according to the recognition results of the first to third 2D image parts Azimuth. The brain tissue recognition module is connected to the transmission module to receive the corrected 3D brain image to be determined, and based on the different voxel characteristics of the white matter, gray matter, and spinal fluid on the brain image, the corrected 3D brain to be determined The image captures a white matter image portion to be determined, a gray matter image portion to be determined, and a spinal fluid image portion to be determined. The processing module is connected to the storage module, the transmission module, and the brain tissue recognition module, and uses the first to third judgment models stored in the storage module to transfer the waiting data from the brain tissue recognition module The determined white matter image portion, the gray matter image portion to be determined, and the spinal fluid image portion to be determined are respectively fed into the input layer of the first to third determination models after dimensionality reduction processing, through the features of the first to third determination models After the calculation of the selection layer and the hidden layer, the first to third judgment results are respectively generated in the output layers of the first to third judgment models, and an evaluation result is generated according to the first to third judgment results and passed through the transmission model. The group outputs the evaluation result to the outside.

The effect of the present invention is to perform azimuth correction processing on each of the 3D brain images used to train the first to third judgment models and the 3D brain images to be judged, thereby effectively removing the head movement caused Interference and help establish accurate judgments Models, and accurate brain image determination; In addition, 3D brain images of normal persons without neuropsychiatric diseases and patients with neuropsychiatric diseases are used to perform models corresponding to different brain tissues such as white matter, gray matter, and spinal fluid. Therefore, the established first to third judgment models not only have relatively high sensitivity for 3D brain images of patients with neuropsychiatric diseases, but also have relatively high sensitivity for 3D brain images of normal persons without neuropsychiatric diseases. High specificity (Specificity). In application, the evaluation results including the first to third judgment results can indeed be provided to psychiatrists and used as an effective auxiliary judgment basis for diagnosing neuropsychiatric patients.

100: Neuropsychiatric Disease Assessment System

10: Determine the platform

1: Storage module

2: Transmission module

3: Azimuth correction module

4: Brain tissue recognition module

5: Processing module

20: Use the terminal

200: Communication network

S301~S307: steps

S801~S807: steps

Other features and effects of the present invention will be clearly presented in the embodiments with reference to the drawings, in which: FIG. 1 is a block diagram exemplarily showing the brain imaging-based neuropsychiatric disease assessment system of an embodiment of the present invention Architecture; Figure 2 is a block diagram exemplarily depicting the architecture of a determination platform of this embodiment. Figure 3 is a flow chart exemplarily illustrating the first to third stored in a storage module of the determination platform How the judgment model is established; Figures 4 to 6 respectively show examples of the first to third 2D images corrected by the azimuth correction module of the judgment platform; Figure 7 is a schematic diagram illustrating this exemplarily The first/second/third created by the embodiment The architecture of the judgment model; and FIG. 8 is a flowchart illustrating how the embodiment performs neuropsychiatric assessment processing on a 3D brain image to be judged.

Before the present invention is described in detail, it should be noted that in the following description, similar elements are denoted by the same numbers.

Referring to FIG. 1, the neuropsychiatric disease assessment system 100 based on brain imaging of the illustrated embodiment of the present invention is mainly used to determine whether a 3-dimentional brain image has abnormalities associated with neuropsychiatric diseases. Brain area. In this embodiment, the to-be-determined 3D brain image may be, for example, a combination of multiple 2D (2-dimentional) MRI images obtained by scanning different sections of the human brain in the axial direction, such as by a magnetic resonance imaging (MRI) instrument. The 3D MRI image is not limited to this example. In other embodiments, it may also be a 3D computerized tomography (CT) image. The neuropsychiatric disease assessment system 100 includes a judgment platform 10 and a terminal 20. The determination platform 10 can be implemented by, for example, a computer system (not shown), and the user terminal 20 can be implemented by, for example, a computer device (such as a desktop computer, a notebook computer, a tablet computer, a smart phone, etc.), and It communicates with the determination platform 10 via a communication network 200 (such as the Internet). In actual use, the judgment platform 10 can be regarded as a cloud server for providing neuropsychiatric disease assessment services. In order to provide this neuropsychiatric disease evaluation service for evaluation requests made by other electronic devices similar to the user terminal 20 (as indicated by the dotted line in FIG. 1).

Referring to FIG. 2, the judgment platform includes, for example, a storage module 1, a transmission module connected to the communication network 200 and used for data transmission 2, an azimuth correction module 3 connected to the transmission module 2, and an azimuth correction module 3 connected to the transmission module 2. The brain tissue recognition module 4 of the azimuth correction module 3, and a processing module 5 connected to the storage module 1, the transmission module 2 and the brain tissue recognition module 4. The operations and functions of the azimuth correction module 3, the brain tissue recognition module 4, and the processing module 5 will be described below.

In this embodiment, before use, the storage module 1 pre-stores a first judgment model, a second judgment model, and a third judgment model corresponding to three types of brain tissues, namely, white matter, gray matter, and spinal fluid. Hereinafter, referring to FIGS. 2 and 3, we will further illustrate in detail how to use the determination platform 10 (or a computer system similar to the determination platform 10) to execute a modeling program to obtain the first to third determination models. The modeling procedure includes the following steps S301~S307.

In step S301, the transmission module 2 obtains a plurality of (for example, 250, but not limited to this example) 3D brain images corresponding to a plurality of normal persons without neuropsychiatric diseases, and a plurality of 3D brain images from outside. (For example, 250, but not limited to this example) 3D brain images corresponding to multiple patients suffering from neuropsychiatric diseases.

Then, in step S302, the azimuth correction module 3 performs azimuth correction processing for each 3D brain image obtained by the transmission module 2. More specifically, In the azimuth correction processing for each 3D brain image, the azimuth correction module 3 first extracts the first to third 2D image parts orthogonal to each other and intersecting at the center of the 3D brain image from the 3D brain image . In this embodiment, for example, the first 2D image portion is a section of the 3D brain image perpendicular to the coronal, as shown in FIG. 4; the second 2D image portion is a section of the 3D brain image The section perpendicular to the sagittal direction, as shown in FIG. 5; and the third 2D image part is an axial section of the 3D brain image, as shown in FIG. 6. Then, the azimuth correction module 3 uses a deep learning algorithm to identify the first to third 2D image parts, so as to correct the center of the 3D brain image according to the identification results of the first to third 2D image parts 3D azimuth. It is worth noting that the deep learning algorithm uses, for example, a Convolutional Neural Network (CNN) regression model to perform image recognition and image comparison processing, but it is not limited to this example. For example, the convolutional neural network regression model used in this embodiment can be planned to have 2 to 4 convolution layers, a full connected layer, and a regression layer. , But not limited to this example, so that the overall calibration process can be consistent. More specifically, in this embodiment, the azimuth correction module 3 uses the deep learning algorithm to determine whether each of the first 2D image portion and the third 2D image portion is approximately symmetrical to each other on the left and right sides. , And identify the two reference image regions contained in the second 2D image portion corresponding to two predetermined brain features, and determine whether the positions of the reference image regions are on a horizontal line. In this In an embodiment, the two predetermined brain features are, for example, Anterior Commissure and Posterior Commissure, respectively, and the two reference image areas are, for example, the first reference image area 51 representing the Anterior Commissure, and The second reference image area 52 representing the post-combination is shown in FIG. 5. Therefore, when the azimuth angle determination module 3 determines that the first 2D image portion is asymmetrical on the left and right sides, it calibrates the first 2D image portion by rotating the first 2D image portion relative to the coronal direction until it is calibrated. The latter first 2D image portion is approximately symmetrical to each other on the left and right sides (see FIG. 4), and the total rotation angle relative to the coronal direction is taken as the corrected angle of the corrected 3D azimuth in the coronal direction. When the azimuth determining module 3 determines that the positions of the reference image areas 51, 52 are not located on a horizontal line, the second 2D image portion is calibrated by rotating the second 2D image portion with respect to the sagittal direction until the The positions of the reference image areas 51, 52 of the calibrated second 2D image portion are located on a horizontal line (see Figure 5), and the total rotation angle relative to the sagittal direction is taken as the corrected 3D azimuth angle in the The correction angle of the sagittal direction. When the azimuth determining module 3 determines that the third 2D image portion is asymmetric on the left and right sides, it calibrates the third 2D image portion by rotating the third 2D image portion relative to the axial direction until the calibrated The third 2D image portion is approximately symmetrical to each other on the left and right sides (see FIG. 6), and the total rotation angle relative to the axial direction is used as the corrected angle of the 3D azimuth angle in the axial direction. With this azimuth correction processing, the 3D brain image (during shooting) can be effectively removed from the interference caused by, for example, head movement. Therefore, the azimuth correction module 3 transmits all the corrected 3D brain images to the brain tissue recognition module 4.

After that, in step S303, the brain tissue recognition module 4 uses the white matter, gray matter, spinal fluid, and other brain tissues in the brain image for each 3D brain image from the azimuth correction module 3 (the azimuth angle has been corrected). Perform a brain tissue recognition process on different Intensity of Voxel distribution characteristics, so as to extract a white matter image part, a gray matter image part and a spinal fluid image part from the 3D brain image, each of them It is also a 3D image. For example, the voxel value distribution range of gray matter, white matter, and spinal fluid can be based on the histogram of the number of voxel (Number of Voxel) versus the voxel value (Intensity of Voxel) obtained by analyzing the 3D brain image (i.e. , The relationship between the value of each voxel and the number of voxels), and is determined by a norm composed of three Gaussian distributions. In this embodiment, the voxel value distribution range of white matter is, for example, 7500-22500; the voxel value distribution range of gray matter is, for example, 5500-20000; and the voxel value distribution range of spinal fluid is, for example, 0-13000, but not here. limit. Therefore, the brain tissue recognition module 4 transmits multiple white matter image parts, multiple gray matter image parts, and multiple spinal fluid image parts captured from all 3D brain images to the processing module 5.

Then, in step S304, in order to reduce the complexity of subsequent calculations, the processing module 5 performs a process for the white matter image parts, the gray matter image parts, and the spinal fluid image parts from the brain tissue recognition module 4 Dimensionality reduction processing. More specifically, since the white matter image parts, the gray matter image parts, and the spinal fluid image parts all have a 3D vector data form, the processing module 5, for example, converts each white matter/gray The 3D vector data form of the qualitative/spinal fluid image part is transformed into a 1D vector data form for subsequent processing.

After that, the processing module 5 uses the deep neural network to perform a deep neural network based on all white matter image parts after dimensionality reduction (for example, 250 white matter image parts corresponding to normal persons, and 250 white matter image parts corresponding to neuropsychiatric patients). The algorithm self-learns and trains to obtain the first judgment model (step S305). In particular, in the actual operation of step S305, the processing module 5 can first use, for example, 200 white matter image parts corresponding to normal persons and 200 white matter image parts corresponding to neuropsychiatric patients together as learning and training. The data part to establish the first judgment model. As shown in FIG. 7, the first decision model includes an input layer, a feature selection layer related to abnormal brain regions of patients with neuropsychiatric diseases, multiple hidden layers, and an output layer. Then, the remaining 50 white matter image parts corresponding to the normal person and the 50 white matter image parts corresponding to the neuropsychiatric patients are used together as the data part for verifying or testing the first judgment model. In other words, if an erroneous judgment occurs (that is, the verification is unsuccessful), for example, when the data fed into the input layer corresponds to the normal white matter image portion, but the output layer generates, for example, an indication that "has one or more The result of determining the abnormal brain area and its location", or when the data fed into the input layer corresponds to the white matter image part of a person with neuropsychiatric disease, and the output layer indicates, for example, "no abnormal brain In this case, the relevant parameters in the feature selection layer can be further adjusted according to the verification result, so as to improve the effect of the first judgment model on neuropsychiatric disorders. Sensitivity of the white matter image of the patient, and specificity of the white matter image of the normal person.

At the same time, similar to the operation of step S305, the processing module 5 calculates all gray matter image parts after dimensionality reduction (for example, 250 gray matter image parts corresponding to normal persons, and 250 gray matter image parts corresponding to neuropsychiatric patients). ), using the same deep neural network algorithm to self-learn and train to obtain the second judgment model (step S306). The second decision model has an architecture similar to the first decision model (see Figure 7). The actual operation of step S306 can be completely similar to the actual operation of step S305 described above, and will not be repeated here. Similarly, the second judgment model obtained by the processing module 5 also has relatively high sensitivity and specificity.

At the same time, similar to the operation of step S305, the processing module 5 is based on all spinal fluid image parts after dimensionality reduction (for example, 250 spinal fluid image parts corresponding to a normal person, and 250 spinal fluid image parts corresponding to a neuropsychiatric patient). Spinal fluid image part), using the same deep neural network algorithm to self-learn and train to obtain the third judgment model (step S307). Similarly, the third decision model has an architecture similar to the first decision model (see FIG. 7). In the actual operation of step S307, it can also be completely similar to the actual operation described in step S305, which will not be repeated here. Similarly, the third judgment model obtained by the processing module 5 also has relatively high sensitivity and specificity.

Hereinafter, referring to Figs. 1, 2 and 8 to explain how the neuropsychiatric disease assessment system 100 performs a 3D brain image of a person to be assessed (such as a patient). Perform neuropsychiatric evaluation and treatment. This neuropsychiatric disease evaluation process includes the following steps S801 to S807.

First, in step S801, the user terminal 20 transmits an evaluation request containing the 3D brain image to be determined to the transmission module 2 of the determination platform 10 via the communication network 200.

Then, in step S802, when the transmission module 2 receives the evaluation request, the azimuth correction module 3 receives the to-be-determined 3D brain image, and performs the above-mentioned step S302 (FIG. 3) for the to-be-determined 3D image. ) The azimuth correction processing described above is to correct the 3D azimuth of the center of the 3D brain image to be determined, and send the corrected 3D brain image to be determined to the brain tissue recognition module 4.

After that, similar to the above-mentioned step S303 (FIG. 3), in step S803, the brain tissue recognition module 4 responds to the to-be-determined 3D brain image from the azimuth correction module 3 (its azimuth angle has been corrected) based on the white matter According to the different voxel characteristics of gray matter and spinal fluid on brain images, a brain tissue recognition process is performed, so that the 3D brain image to be determined after self-correction is extracted, a white matter image part to be determined, a gray matter image part to be determined, and A spinal fluid image portion to be determined, and the white matter image portion to be determined, the gray matter image portion to be determined, and the spinal fluid image portion to be determined are transmitted to the processing module 5.

Then, similar to the above-mentioned step S304 (FIG. 3), in step S804, the processing module S804 treats the white matter image portion to be determined, the gray matter image portion to be determined and the gray matter image portion to be determined from the brain tissue recognition module 4 Spinal fluid imaging is partially performed Dimension processing.

After that, in step S805, the processing module 5 uses the first to third determination models stored in the storage module 1 to convert the white matter image portion to be determined and the gray matter image portion to be determined in the form of 1D vector data And the spinal fluid image part to be judged are respectively fed into the input layers of the first to third judgment models, and after the calculation of the feature selection layer and hidden layer of the first to third neuropsychiatric disease judgment models, in the first The output layers of the first to third decision models generate first to third decision results, respectively. For example, if the first determination model is calculated to determine that the white matter image portion to be determined does not have any abnormal brain regions, the first determination result will indicate, for example, "FALSE1" (which represents that the white matter image portion to be determined does not have any abnormal brain regions). Have any abnormal brain areas associated with neuropsychiatric diseases); or if the first determination model is calculated to determine that the white matter image portion to be determined has one or more abnormal brain areas associated with neuropsychiatric diseases, then The first determination result will indicate, for example, "TRUE1" (which indicates that the white matter image portion to be determined has an abnormal brain area associated with a neuropsychiatric disease), and location data of the abnormal brain area(s). But not limited to this example. Similarly, the second determination result can indicate one of the following two situations: one is to indicate that the gray matter image portion to be determined has one or more abnormal brain regions associated with neuropsychiatric diseases (for example, "TRUE2 "To indicate) and the location data of the abnormal brain area(s); and another case indicates that the gray matter image portion to be determined does not have any abnormal brain areas associated with neuropsychiatric diseases (for example, "FALSE2" Express). In the same way, the third judgment result can indicate One of the following two situations: One is to indicate that the spinal fluid image portion to be determined has one or more abnormal brain areas associated with neuropsychiatric diseases (for example, represented by "TRUE3") and the (etc.) Location data of abnormal brain regions; and another situation indicates that the spinal fluid image portion to be determined does not have any abnormal brain regions associated with neuropsychiatric diseases (for example, represented by "FALSE3").

Then, in step S806, the processing module 5 generates an evaluation result about abnormal brain regions according to the first to third determination results, and responds to the evaluation request with the evaluation result as an evaluation response via the transmission module 2 is transmitted to the user terminal 20.

Finally, when the user terminal 20 receives the evaluation result from the determination platform 10, the user terminal 20 displays the evaluation result so that the user (such as a psychiatrist) can quickly learn that the person to be evaluated is in white matter and gray matter. And the determination status of spinal fluid (step S807). It is worth mentioning that, in order to make it easier to interpret the judgment output, especially when the evaluation result contains one or more pieces of position data, the user terminal 20 can use the pieces of position data contained in the evaluation result, for example, in the The 3D brain image to be determined is marked with a specific mark on the abnormal brain area corresponding to the position data (e.g.), and the 3D brain image to be determined and the specific mark added are displayed at the same time, so that the user can use The displayed specific markers quickly get the location of the abnormal brain area.

In summary, the azimuth correction processing is performed on each 3D brain image used to train the first to third judgment models and the 3D brain image to be judged, thereby It can effectively remove the interference caused by head movement and help to establish accurate judgment models and accurate brain image judgments; in addition, because the 3D brains of normal persons without neuropsychiatric diseases and patients with neuropsychiatric diseases are used Images are used to perform model training corresponding to different brain tissues such as white matter, gray matter, and spinal fluid. Therefore, the first to third judgment models established are not only relatively sensitive to 3D brain images of patients with neuropsychiatric diseases, but also The 3D brain images of normal persons without neuropsychiatric diseases also have relatively high specificity. In application, the evaluation results including the first to third judgment results can indeed be provided to psychiatrists and used as an effective auxiliary judgment basis for diagnosing neuropsychiatric patients. Therefore, it can indeed achieve the purpose of the invention.

However, the above are only examples of the present invention. When the scope of implementation of the present invention cannot be limited by this, all simple equivalent changes and modifications made in accordance with the scope of the patent application of the present invention and the content of the patent specification still belong to This invention patent covers the scope.

S301-S307...Steps

Claims (9)

A neuropsychiatric disease assessment method based on brain images is implemented by a computer system and includes the following steps: (A) Obtain multiple 3D brain images corresponding to multiple normal persons without neuropsychiatric diseases, and multiple separate Corresponding to the 3D brain images of multiple patients with neuropsychiatric diseases; (B) For each 3D brain image obtained in step (A), perform azimuth correction processing, in which a deep learning algorithm is used to identify The first to third 2D image parts captured from the 3D brain image and orthogonal to each other and intersecting at the center of the 3D brain image, so as to correct the 3D brain image according to the recognition results of the first to third 2D image parts The 3D azimuth angle of the center; (C) For each corrected 3D brain image, based on the different voxel characteristics of white matter, gray matter and spinal fluid on the brain image, perform a brain tissue recognition process to extract the 3D brain image, Extract a white matter image part, a gray matter image part and a spinal fluid image part; (D) All white matter image parts, all gray matter image parts and all spinal fluid images obtained according to step (C) and after dimensionality reduction processing Part, using deep neural network algorithm self-learning training to obtain the first to third judgment models corresponding to gray matter, white matter, and spinal fluid. Each of the first to third judgment models includes an input layer, one and A feature selection layer, multiple hidden layers, and an output layer related to abnormal brain regions of patients with neuropsychiatric diseases; (E) When a 3D brain image to be determined is received, the 3D brain image to be determined is executed in sequence Finish the azimuth correction processing and the brain tissue recognition Processing to obtain a white matter image portion to be determined, a gray matter image portion to be determined, and a spinal fluid image portion to be determined corresponding to the 3D brain image to be determined; (F) the white matter image to be determined after dimensionality reduction processing Part, the gray matter image part to be judged and the spinal fluid image part to be judged are fed into the input layer of the first to third judgment models, through the feature selection layer and the hidden layers of the first to third judgment models After the calculation of, first to third judgment results are respectively generated in the output layers of the first to third judgment models; and (G) an evaluation result related to abnormal brain regions is generated according to the first to third judgment results. The method for evaluating neuropsychiatric diseases based on brain images according to claim 1, wherein, in step (B), for each 3D brain image: the first to third 2D images captured from the 3D brain image The parts are three slices of the 3D brain image that are perpendicular to the coronal, sagittal, and axial directions; and in the azimuth correction process, the deep learning algorithm is used to determine the first 2D image part and the second Whether each of the three 2D image parts is roughly symmetrical to each other on the left and right sides, and identify the two reference image regions contained in the second 2D image part corresponding to two predetermined brain features, and determine the reference images Whether the location of the zone is located on a horizontal line, when it is determined that the first 2D image portion is asymmetrical on the left and right sides, the first 2D image portion is calibrated by rotating the first 2D image portion relative to the coronal direction until it is calibrated After the first 2D image part is on the left and right The two sides are approximately symmetrical to each other, and the total rotation angle relative to the coronal direction is taken as the corrected angle of the 3D azimuth angle in the coronal direction after correction. When it is determined that the locations of the reference image areas are not on a horizontal line, Calibrate the second 2D image portion by rotating the second 2D image portion relative to the sagittal direction until the positions of the reference image regions of the second 2D image portion after calibration are located on a horizontal line, and will be relative to the The total rotation angle of the sagittal direction is taken as the corrected angle of the corrected 3D azimuth in the sagittal direction, and when it is determined that the third 2D image portion is asymmetric on the left and right sides, the third 2D image is rotated relative to the axial direction. Partially calibrate the third 2D image part until the calibrated third 2D image part is roughly symmetrical to each other on the left and right sides, and use the total rotation angle relative to the axis as the corrected 3D azimuth on the axis Correction angle to the direction. The method for evaluating neuropsychiatric diseases based on brain images according to claim 1, wherein, in step (B), the deep learning algorithm uses a convolutional neural network regression model to perform image recognition and image comparison processing. The method for evaluating neuropsychiatric diseases based on brain imaging according to claim 1, wherein, in step (F), each of the first to third determination results indicates the portion of the white matter image to be determined and the portion of the white matter image to be determined One of the gray matter image portion and the spinal fluid image portion to be determined has at least one abnormal brain area associated with neuropsychiatric disease and location data of the at least one abnormal brain area, or indicates that the corresponding image portion does not have any Abnormal brain areas associated with neuropsychiatric diseases. A neuropsychiatric disease assessment system based on brain imaging, including: A judgment platform, including a storage module, storing the first to third judgment models as described in claim 1, a transmission module, adapted to receive a 3D brain image to be judged, an azimuth correction module, connected The transmission module receives the 3D brain image to be determined, and extracts first to third 2D image parts orthogonal to each other and intersecting at the center of the 3D brain image to be determined from the 3D brain image to be determined, using a deep learning The algorithm performs image recognition of the first to third 2D image parts, and corrects the 3D azimuth angle of the center of the 3D brain image to be determined according to the recognition results of the first to third 2D image parts, a brain tissue recognition The module is connected to the transmission module to receive the corrected 3D brain image to be determined, and based on the different voxel characteristics of white matter, gray matter and spinal fluid on the brain image, self-corrected 3D brain image to be determined is extracted A white matter image portion to be determined, a gray matter image portion to be determined, and a spinal fluid image portion to be determined, and a processing module connected to the storage module, the transmission module and the brain tissue recognition module, and use the storage The first to third judgment models stored in the module feed the white matter image part to be judged, the gray matter image part to be judged and the spinal fluid image part to be judged from the brain tissue recognition module respectively after dimensionality reduction processing. Enter the input layers of the first to third decision models, and after the calculation of the feature selection layers and hidden layers of the first to third decision models, first to third decision models are generated in the output layers of the first to third decision models, respectively. The third determination result, and an evaluation result is generated according to the first to third determination results, and the evaluation result is externally output through the transmission module. The neuropsychiatric disease assessment system based on brain images according to claim 5, wherein: the first to third 2D image parts captured by the azimuth correction module are three respectively vertical parts of the 3D brain image to be determined Cut planes in the coronal, sagittal, and axial directions; and the azimuth correction module uses the deep learning algorithm to determine whether each of the first 2D image portion and the third 2D image portion are substantially opposite to each other on the left and right sides It is symmetrical, and the two reference image regions contained in the second 2D image portion corresponding to two predetermined brain features are identified, and it is determined whether the positions of the reference image regions are located on a horizontal line. When the 2D image portion is asymmetric on the left and right sides, the first 2D image portion is calibrated by rotating the first 2D image portion relative to the coronal direction until the calibrated first 2D image portion is approximately symmetrical to each other on the left and right sides. , And regard the total rotation angle relative to the coronal direction as the corrected angle of the corrected 3D azimuth in the coronal direction. When it is determined that the positions of the reference image areas are not located on a horizontal line, it is relative to the sagittal direction Rotate the second 2D image portion to calibrate the second 2D image portion until the positions of the reference image regions of the second 2D image portion after calibration are located on a horizontal line, and the total rotation angle relative to the sagittal direction As the corrected angle of the corrected 3D azimuth in the sagittal direction, and when it is determined that the third 2D image portion is asymmetric on the left and right sides, the third 2D image portion is calibrated by rotating the third 2D image portion relative to the axial direction. The third 2D image part until the calibrated third 2D image part is on the left and right. The sides are approximately symmetrical to each other, and the total rotation angle relative to the axial direction is taken as the corrected angle of the 3D azimuth angle in the axial direction after correction. The neuropsychiatric disease assessment system based on brain images of claim 5, wherein the deep learning algorithm uses a convolutional neural network model to perform image recognition and image comparison processing. The neuropsychiatric disease assessment system based on brain imaging according to claim 5, wherein each of the first to third determination results indicates the white matter image portion to be determined, the gray matter image portion to be determined, and the spinal cord to be determined One of the corresponding image parts of the liquid image part has at least one abnormal brain area associated with neuropsychiatric disease and location data of the at least one abnormal brain area, or indicates that the corresponding image part does not have any abnormality associated with neuropsychiatric disease Brain area. The neuropsychiatric disease assessment system based on brain images according to claim 5, further comprising: a user terminal that communicates with the transmission module via a communication network to transmit the to-be-determined 3D brain image to the transmission module and Receive the evaluation result from the transmission module; wherein, when the evaluation result contains one or more pieces of position data, the terminal uses the position data (e.g.) contained in the evaluation result to perform the determination on the 3D brain The image displays the to-be-determined 3D brain image and the added specific mark at the same time in a manner that a specific mark is added to the abnormal brain area corresponding to the position data(s).

Images